An Introduction to Spotify’s Genres of the Future

Spotify’s Genres of the Future - A Study on Spotify’s Genre Classification

Spotify has several curious genres within its database, including some which are characterized by the word future. Namely, the following selection:

It is definitely not clear if these genres are connected by something more than just their name. We’ll explore why these genres are called the way the are, and if a (strong) connection might actually tie these genres together. This will consist of exploring the aspects of the music within each genre first, after which we will properly compare the 5 genres by defining a suitable corpus and providing insightful visualizations of the audio analysis data provided by the Spotify API, using some features provided by the compmus R package.

To begin our journey through the future, we take a look at the ability of statistical classifiers to differentiate between the five genres, using a k-neighbours classifier with variables extracted from a random forest classifying tactic. We will find that the Spotify API itself might not be suitable to finding a satisfying answer to our question, although it will tell us exactly how the genres are in fact quite different from each other.

Lets first take a step back however, and get acquainted with the genres and their surface level features.





Some sound samples from the different genres (in fact, these playlist together make up the entire corpus).

A Future of Sounds, Rhythms and Timbres - A Classification by Spotify’s API features

Random Forest Classification

To start the classification process, we begin by applying random forest classification to our corpus. We immediately notice that most important variables in this model are not at all what we suspected in our preliminary analysis. The model the duration of the song and favors characteristics of the sound such as cepstro coefficient 1, which indicated loudness, cepstro coefficient 2, which indicates the brightness of the sound, and loudness again. Indeed, all these properties fluctuate from genre to genres, but they are not very insightful for our research question.

Nevertheless, we will use the top five variables for a k-neighbour classification from which the confusion matrix can be found below.

For our research this is of course useful information to know, but we are more interested in where these genres are similar. Aside from the none key-based properties, we can see that the forest classifier found acousticness and liveness to be least important for classification. This leads us to a right path: the future is filled with low acousticness and liveness.


Naive K-neighbour Classification

In the ad hoc analysis we suspected valence and energy might be an important factors to differentiate between the genres. This was the motivation for the following k-neighbour classification and its respective confusion matrix. Apparently however, the classification does not do superbly well.

Random Forest Based K-neighbour Classification

We thereafter look at a k-neighbour classification using the top variables found during the random forest classification, which does significantly better than the naive model. Here we find some interesting results concerning which genre combinations the model is better at finding the correct classification. To begin, it achieves high accuracy for futurepop, indicating its unique nature between the five genres.

Secondly, the classifier seems to struggle a lot when differentiating between future funk and kawaii future bass. Due to the reasons above, this does not come as a huge surprise. Between the other three genres though, the classifier does a great job, further indicating their differences from the group.

Additionally, future garage and future ambient are difficult to seperate as well, and is seems like future garage is confused a lot with future funk as well! Thus clearly, these genres have a lot in common as well. There is no confusion when differentiating these with kawaii future bass or futurepop however, further indicating their respective differences.


A Short Ad Hoc Examination of the Genres

Future Funk and Kawaii Bass

From exploring the genres and listening to some excerpts of their music, we can quickly suspect that the future description might mean something entirely different per genre. In the case of Future Funk and Kawaii Future Bass, it seems to refer to a futuristic pop-y sound, with a distinct electronic feel, although real (sampled) instruments and some vocals (often with light-hearted lyrics) now and then are common too, especially in Future Funk. Future Kawaii Bass sets itself apart from Future Funk with extensive use of chiptune sounds and strong upbeat rhythms.

Futurepop

Futurepop lies on the complete other end of the spectrum however, with dark low synths and rhythms characterizing its sound. These support raw, unedited vocals with lyrics I could only describe as “desperate”. I believe that if the Future in its name refers to anything else than the extensive usage of synthesized sounds, it would refer to a distinctively dystopian and dark impression of the future.

Future Garage and Future Ambient

Future Garage and Future Ambient could both in a certain way be described as minimalistic. They use little instruments, which often repeat short musical fragments throughout the whole song in a relatively low tempo. Strong drum grooves give structure to the songs, in which generally no distinctive melody or vocal part is present, especially in Future Ambient. There, slow reverberized synth sounds and sweeps give color to tracks. In this regard, Future Garage is certainly different. It features these synths too, but also contains a lot more acoustic “real” instruments, albeit often sampled, and the occasional “smooth” melody line. For both genres, the Future seems simply to be an indication of the electronic style and heavy usage of sampling.

The Creation of the Corpus

The collections of songs we’ll analyze, our corpus, will naturally be a selection of songs from artist from each of these genres. Large names such as Snail’s House (who has in fact been credited to be the pioneer of Kawaii Future Bass, but is also listed as a Future Funk artist) will be of particular interest and multiple songs will be included, but some genres such as future ambient consist of mostly smaller artist, where a smaller and more varied selection of songs from different artist is more appropriate.

It could also be helpful to compare some of the outliers of the genre, but care must be taken to ensure those artist would actually belong to the genre, instead of their appearance being based solely on Spotify’s automatic assignment. If not, it will be more useful to exclude them from the research, as they would negatively interfere with making valid comparisons between the genres.

I ended up putting together a playlist for every genre based on playlists provided by the Every Noise at Once website (http://everynoise.com/), each around 50 tracks long. The Every Noise at Once website is a large compilation of all genres on Spotify, and should give a representative selection of songs for each genres, at least according to Spotify’s own genre classification algorithms.

The song count for the five genres
Genre Song Count
Future Ambient 50
Future Funk 49
Future Garage 50
Futurepop 50
Kawaii Future Bass 49

Naive K-neighbour Classification

Random Forest Based K-neighbour Classification

Applying Clustering to the Genres


Description and Results

Clustering using a randomly selected subset of the corpus shows that the five genres in fact cluster into just three large clusters, although it is difficult to categorize these songs manually based on listening and inspection. When one listens the songs from each respective cluster in the dendogram on the left however, one can hear the similarities somewhat. Once again, the track duration seems to come back prominentely, further indicating its importance.

From this graph, it is clear that a lot of overlap exists between the genres. However, we can not begin to make claims about the uniformity among the future genres. In fact, we emphasize some of the larger differences when looking at a valence plot in the next section, and chromagrams, cepstrograms and chordograms in section thereafter.

A Deeper Investigation into he Valence and Energy of the Genres


Description

An important finding of our preliminary research was the large deviations of musical style between the genres. To me, valence seemed to be a logical choice for comparison, as I felt like the genres sounded most distict in this regard (even though the classification process suggested otherwise). The graph shows the average valence of the songs from each genre, as well as the value of each individual song in the form of a scatter plot. Additionally, the colour displays the energy value of the tracks, another feature I felt provides contrast between the genres. Lastly, the size of each point gives an indication of the liveness of the track, as provided by Spotify.

Results

Surprisingly, we can see Futurepop actually has a somewhat high average, even higher than Kawaii Future Bass, something I think one would not concluded based on listening to the selections. As expected, Future Funk and Kawaii Future Bass do in general seem to be assigned higher valence values than Future Ambient and Future Garage.

Chroma, Cepstro and Chord Features




Description

Here we can see a chromagram, cepstrogram and chordogram of a single somewhat representative song for every genre. The songs are picked by hand, which ensures a totally subjective view of every genre, but I think it can be a useful tool for analysis nonetheless. In no particular order, these songs are:

  • Forever (Original Club Mix) - Bruderschaft (Futurepop)
  • Ji-Eun’s Sunset - City Girl (Future Funk)
  • Hot Milk - Snail’s House (Kawaii Future Bass)
  • King Bromeliad - Floating Points (Future Garage)
  • Freefloat - Aural Float (Future Ambient)

Each song is compared using the same norm to allow for fair comparisons.


Results

The first thing one may notice is the differences in repetitions between the genres. The Future Garage song especially has a very homogeneous structure judging from the three c-grams. Future Ambient is similar, but seems to have two different sections which both show a lot of individual repetition.

In the chordograms, we see that future funk and future kawaii bass both seem to be more varied with their chord changes in comparison to the other genres. We can also observe that, along with futurepop, these songs contain a reoccuring structure in the c-grams which the other two genres do not. This could, for example be due to the presence or non-presence of the common chorus-verse structure.

Furthermore, the cepstro features show more differences. The Future Ambient song shows a dynamic progression, where the different coefficients change in prevalance throughout the song. Kawaii Future Bass and Futurepop show the ‘chorus-structure’ we have discussed, while the remaining two genres seem to stay somewhat consistent throughout, only changing in a small section within the piece. For all songs, the first coefficient, loudness, stays very consistent throughout the song, but each are at different levels. Loudness could be an effect of the recording/mastering process for each genre, which this data does suggest could be different.

What makes the ‘future’, the future? - Conclusion



What makes the ‘future’, the future?

We have found a selection of music which is similar within in many ways, but always only within subsets of the whole collection.
Appart from the lack of acousticness and liveness these genres share - a common feature of almost all modern studio recorded music -,
there seem to be no characteristics which these genres of the future all share.

Future Funk and Future Kawaii Bass have a lot of overlap, but just with the two of them. So do Future Ambient and Future Garage,
prefering long minimalistic and repeating pieces. Alas, the other three genres do not. Could it be the electronic nature which Kawaii Bass,
Future Pop and Future Ambient have in common? Sadly, Future Funk and Future Garage do not play along. And sadly, so this trend seems to continue.
Additionally, the effectiveness of the classification which we showed in the confusion matrix is further indication of this fact.


The Future is more like an idea than it is a characteristic. It is an idea that means different things to different artists.
As a label, its applicibility to the give genres can be argued over, but overall, it is something one can interpret as they want,
to indicate whatever they want.

As for Spotify, they seem to be happy to define it as just that.